Trust Is the Only Layer You Can't Orchestrate
Written by Robin Caller on June 24, 2026
Two things happened this month that tell you where the data stack is heading. On 1 June, Fivetran and dbt Labs closed their merger and planted a flag: they are building, in their own phrase, the data infrastructure for trusted AI agents. Nine days later Adobe put CX Enterprise Coworker into general availability, a “super agent” whose job is to run a company’s other agents. Ingestion and transformation and activation and the agents themselves, all of it folding into one coordinating layer. Orchestration is the centre of gravity now, and I think that is roughly right. We are building toward it ourselves.
One word in the launch copy keeps snagging me, though. Trusted.
Everyone Is Selling "Trusted." Almost Nobody Defines It.
Read the copy and the word is everywhere. Trusted AI agents, trusted data, a foundation you can rely on. It is the load-bearing claim of the whole 2026 pitch, and hardly anyone stops to ask what it is supposed to mean.
So ask. What does orchestration actually do to your data? It moves records between systems, reshapes them for wherever they are going, pushes them out to the CRM and the ad platforms, and tries to get the timing right. That is real engineering and worth paying for, but notice that all of it operates on records as they already are. None of it reaches back and makes one of them true.
Coordination is not correction. A faster, cleaner pipe to more places does nothing for the truth of whatever is travelling down it, and if the thing you put in at the top is wrong, you have just bought yourself a very efficient way of being wrong in a lot more places.
Why You Cannot Orchestrate Your Way to Trust
The reason is structural. Orchestration only ever works on records that already exist, which means that by the time anything is being coordinated the record was captured a while back and whatever was true or false about it is already baked in. The moment that decided it happened upstream, well before the orchestration layer ever saw the thing.
The better platforms have clocked this and bolted on checks, and to be fair it is a real improvement: pipelines now pause to test for anomalies and freshness and schema drift before anything moves. But look at where that check sits. It is mid-pipeline, after capture, once the bad record already exists and has a row of its own. You are inspecting the parcel on the conveyor belt. Whether it should ever have been posted is a question that got skipped about three steps earlier.
None of this is a knock on the tools. It is just the reach of a coordination layer, which only ever begins after the record is already made. Trust is a property of how a record came to exist, and orchestration arrives too late in its life to change that. It can carry trust that is already there, and a good layer should, but it is not the place trust gets created.
The Stakes Changed When the Reader Became a Machine
For years you could half-get-away with this, because the thing at the end of the pipe was a person. A marketer pulled a report, looked at it, used their judgement. A wrong row in a spreadsheet got caught by someone who knew the account. The cost of a bad record was a slightly worse decision, taken next week, by a human who could smell that something was off.
That reader has changed. More and more, the thing standing at the end of your orchestration is an autonomous agent, acting in real time, spending real money, with no instinct of its own. It takes the context it is handed at face value and does not pause to wonder whether the segment is even real. So a bad signal stops being a wasted lead sitting in a list and becomes an instruction, and the smoother your orchestration, the more faithfully that instruction gets carried out. The lag between a wrong record and its consequence used to run in weeks, the time it took a person to notice. With an agent in the loop it runs in minutes.
The mechanism is not abstract, so picture it. A bidding engine like Google’s learns from the conversions you feed it. It does not stop to check that a conversion was real; it takes whatever you have marked as success and goes looking for more people like them. Feed it leads that looked converted but never closed, and it spends your budget chasing an audience that was never going to buy. The instruction was wrong, the machine had no way of knowing, and it spent the money anyway.
And there is a lot of wrong to spend. In our own work, more than a third of the leads we look at are unserviceable by the time anyone tries to act on them, and by unserviceable I mean the mundane stuff: the contact has changed jobs, or it is a duplicate, or the person left the company months ago and nobody updated the record. No fraud, nothing dramatic, just data quietly going off in the way data always has. The difference is that the old world had a person who caught most of it before it did any damage, and an agent with a media budget has no equivalent of that person.
You can see the scale in the wider numbers. When Supermetrics surveyed 435 marketers for its 2026 report, only a third of them said they could activate their data effectively, a fair way behind the share who were happy enough analysing it. The thing they most often blamed was the gap between their analytics and their activation tools, and a good number are already piping first-party data straight back into paid media. So the loop is live in plenty of companies that, pushed on it, would admit they cannot yet act on their data cleanly.
Trust Has to Be Built at the Point of Capture
If trust cannot be added in transit, that leaves one place to build it, which is the point of capture, before the record is even written and long before it becomes a signal an agent acts on. The decision that matters is whether to let it in at all.
We organise around that decision, and the underlying idea is an old one. Rather than fishing the bad apples out of the barrel after the fact, you keep them out of the barrel in the first place. In practice that means a record has to clear three questions before it counts for anything. Is it compliant, meaning lawfully obtained and properly sourced. Is it true, meaning a real person in a role they actually hold today. Does it carry value, meaning the right account and not some stray click. Compliance, truth, value, and the order matters, because there is no sense scoring or routing a record you have not first established is real.
Which is the bit most stacks get backwards. They ingest everything, orchestrate it in real time, and check downstream if they check at all. Turn the sequence around, validate then score then orchestrate, and the orchestration layer finally has something worth coordinating. The catch is that it is never a one-off. Contact data goes off at a rate Landbase puts anywhere between 22 and 70 per cent a year depending on the field, and RevenueBase has business email drifting three or four per cent every month, so the gate at capture has to keep running for as long as new data keeps arriving, which is to say always.
Where LeadScale Is Placing Its Bet
That gap is the whole reason we are building LeadScale as a data orchestration platform rather than something perched above the category passing judgement. We want to be in the orchestration race. The difference is where we start: validation runs at the point of capture, on the Smart Form, before a record is ever allowed to become a signal, so what moves into the orchestration layer has already been checked. The platforms consolidating the stack this year can coordinate data beautifully. What they coordinate is still only worth as much as the truth of it, and that is the ground we have chosen.
Everyone can orchestrate now. The thing that will separate the operations worth trusting is duller than it sounds: whether anyone checked the data was true before the machine started spending money on it. If you fix one thing this year, make it that, and make it happen at the point the data comes in.
